Solution To Land Classification Challenge¶
In [1]:
import numpy as np
import pandas as pd
In [2]:
import matplotlib.pyplot as plt
import seaborn as sns
%matplotlib inline
In [73]:
train_path = '/home/gaurav/Desktop/Social Cops/socialcops_challenge/socialcops_challenge/land_train.csv'
test_path = '/home/gaurav/Desktop/Social Cops/socialcops_challenge/socialcops_challenge/land_test.csv'
data = pd.read_csv(train_path)
test = pd.read_csv(test_path)
In [4]:
print("The first 5 rows of the training dataset\n")
data.head(5)
Out[4]:
In [5]:
print("Basic summary of the training dataset")
data.describe()
Out[5]:
In [6]:
data.info() #to check null values
Clearly there are no missing values in the data!
Plotting the distribution of each column(feature)¶
In [7]:
sns.set(style='whitegrid')
fig, axis = plt.subplots(nrows=1, ncols=6, figsize = (25,4))
for i in range(6):
axis[i].set_title("Distribution Plot {}".format('X'+str(i+1)))
sns.distplot(data['X'+str(i+1)], ax = axis[i])
plt.show()
From the distribution plots we can easily infer that X1,X2 and X3 features are very much correlated as their plots are very similar to each other
In [9]:
sns.lmplot('X1','X2', data=data, hue='target',fit_reg=True, size=5)
sns.lmplot('X1','X3', data=data, hue='target',fit_reg=True, size=5)
sns.lmplot('X2','X3', data=data, hue='target',fit_reg=True, size=5)
Out[9]:
From the plots above, it can be observed how X1, X2, X3 are correlated and the plot between X2 and X3 show very high positive correlation irrespective of class
In [10]:
sns.set(style='whitegrid')
fig, axis = plt.subplots(nrows=1, ncols=6, figsize = (25,4))
for i in range(6):
axis[i].set_title("Distribution Plot {}".format('I'+str(i+1)))
sns.distplot(data['I'+str(i+1)], ax = axis[i])
plt.show()
Similarly I2, I3 and I4 also seem correlated
In [11]:
sns.lmplot('I1','I3', data=data, hue='target',fit_reg=True, size=5)
sns.lmplot('I1','I4', data=data, hue='target',fit_reg=True, size=5)
sns.lmplot('I3','I4', data=data, hue='target',fit_reg=True, size=5)
Out[11]:
Scatter between I1 and I4 show high correlation irrespective of class
In [12]:
sns.pairplot(data, hue="target", size=3)
Out[12]:
Plotting the Correlation HeatMap¶
In [13]:
# Heatmap to see correlation of different feature
f, ax = plt.subplots(figsize=(13,13))
sns.heatmap(data.corr(), annot=True)
plt.show()
Feature Selection¶
Feature importance¶
In [4]:
from sklearn.ensemble import ExtraTreesClassifier
In [5]:
model = ExtraTreesClassifier()
model.fit(data.values[:,:-1],data.values[:,-1])
print(model.feature_importances_)
Feature selection by comparing explained feature variance¶
In [4]:
from sklearn.preprocessing import StandardScaler
# Separating out the features
x = data.iloc[:,:].values
# Separating out the target
y = data.loc[:,['target']].values
# Standardizing the features
x = StandardScaler().fit_transform(x)
In [5]:
from sklearn.decomposition import PCA
pca1 = PCA()
x = pca1.fit_transform(x)
pca1.explained_variance_ratio_
Out[5]:
In [6]:
from sklearn.preprocessing import StandardScaler
# Separating out the features
x = data.iloc[:,:6].values
# Separating out the target
y = data.loc[:,['target']].values
# Standardizing the features
x = StandardScaler().fit_transform(x)
In [7]:
from sklearn.decomposition import PCA
pca2 = PCA()
x = pca2.fit_transform(x)
pca2.explained_variance_ratio_
Out[7]:
In [8]:
from sklearn.preprocessing import StandardScaler
# Separating out the features
x = data.iloc[:,6:-1].values
# Separating out the target
y = data.loc[:,['target']].values
# Standardizing the features
x = StandardScaler().fit_transform(x)
In [9]:
from sklearn.decomposition import PCA
pca3 = PCA()
x = pca3.fit_transform(x)
pca3.explained_variance_ratio_
Out[9]:
Explained Variance
The explained variance tells us how much information (variance) can be attributed to each of the principal components.
Clearly the last componnent or feature in this case contains the least variance, therefore we drop it
In [10]:
data = data.drop(labels='I6',axis=1)
Dealing with Outliers¶
Z-Score Approach
Reference : https://towardsdatascience.com/ways-to-detect-and-remove-the-outliers-404d16608dba
In [11]:
from scipy import stats
import numpy as np
z = np.abs(stats.zscore(data.iloc[:,:-1]))
print(z)
In [12]:
filtered = data[(z < 3).all(axis=1)]
In [13]:
filtered.shape
Out[13]:
In [14]:
data.shape
Out[14]:
Generating pie chart to visualize the number of examples in each target class¶
In [15]:
plt.hist(filtered.iloc[:,-1])
Out[15]:
In [16]:
filtered['target'].value_counts()
Out[16]:
Training and Validation Splitting¶
In [17]:
from sklearn.model_selection import train_test_split
X_train, X_val, y_train, y_val = train_test_split( filtered.iloc[:,:-1],filtered.iloc[:,-1],test_size=0.25, random_state=1)
print(X_train.shape)
print(X_val.shape)
print(y_train.shape)
print(y_val.shape)
Balancing the dataset¶
Using SMOTE - Synthetic Minority OverSampling Method
In [18]:
from imblearn.over_sampling import SMOTE
In [19]:
sampler = SMOTE()
X_train, y_train = sampler.fit_resample(X_train,y_train)
In [20]:
plt.hist(y_train)
Out[20]:
One Hot encoding the categorical data
In [21]:
from keras.utils import np_utils
In [22]:
y_train = np_utils.to_categorical(y_train)
y_val = np_utils.to_categorical(y_val)
print(y_train)
print(y_train.shape)
print(y_val.shape)
In [23]:
y_train = y_train[:,1:]
y_val = y_val[:,1:]
In [24]:
print(y_train.shape)
print(y_val.shape)
Normalizing Data¶
In [25]:
from sklearn.preprocessing import StandardScaler
normalizer=StandardScaler()
X_train=normalizer.fit_transform(X_train)
X_val = normalizer.transform(X_val)
print(X_train.shape)
print(X_val.shape)
Training the Data¶
In [56]:
#Importing Keras Libraries
import keras
from keras.layers import Input,Dense,BatchNormalization, Activation, Dropout, Add
from keras.models import Model
from keras.utils.vis_utils import plot_model
from keras import optimizers
In [64]:
epochs=50
number_of_classes=4
batch_size=150
In [65]:
inpt = Input(shape=(11,), name='input')
x1 = Dense(units=20)(inpt)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x2 = Dense(units=20)(x1)
x2 = BatchNormalization()(x2)
x2 = Activation('relu')(x2)
x2 = Dropout(0.4)(x2)
x3 = Dense(units=20)(x2)
x3 = BatchNormalization()(x3)
x3 = Activation('relu')(x3)
x4 = Add()([x1,x3])
x4 = BatchNormalization()(x4)
x4 = Activation('relu')(x4)
x4 = Dropout(0.2)(x4)
out = Dense(units=number_of_classes,activation='softmax')(x4)
model = Model(inputs=inpt, outputs=out)
In [66]:
model.summary()
In [67]:
checkpoint = keras.callbacks.ModelCheckpoint('weights.{epoch:02d}-{val_loss:.2f}.hdf5', monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1)
adam = optimizers.Adam(lr=0.01, decay=0.0001)
model.compile(optimizer = adam, loss = 'categorical_crossentropy', metrics = ['accuracy'])
output = model.fit(x=X_train,y=y_train,batch_size=batch_size, epochs=epochs,validation_data = (X_val,y_val), callbacks=[checkpoint])
In only 50 epochs we can see that validation loss is less than training loss, hence we can infer that the model is generalizing
In [68]:
accs=output.history['acc']
val_accs=output.history['val_acc']
x_axis=[i+1 for i in range(epochs)]
plt.plot(x_axis,accs)
plt.plot(x_axis,val_accs)
plt.show()
In [69]:
loss=output.history['loss']
val_loss=output.history['val_loss']
x_axis=[i+1 for i in range(epochs)]
plt.plot(x_axis,loss)
plt.plot(x_axis,val_loss)
plt.show()
In [70]:
# Loading the model and Weights of which gave the least validation loss
from keras.models import load_model
model = load_model('/home/gaurav/Desktop/Social Cops/socialcops_challenge/socialcops_challenge/weights.47-0.04.hdf5')
Confusion Matrix
In [71]:
from sklearn.metrics import classification_report
y_val_pred=np.argmax(model.predict(X_val),axis=1)
y_val_pred=y_val_pred+1
y_val_n = np.argmax(y_val,axis=1)
y_val_n+=1
print(classification_report(y_val_n, y_val_pred))
Making predictions¶
In [74]:
test = test.drop('I6',axis=1)
X_test = np.array(test)
X_test = normalizer.transform(X_test)
In [75]:
X_test.shape
Out[75]:
In [76]:
y_pred = np.argmax(model.predict(X_test), axis=1)
In [77]:
print(y_pred)
In [78]:
y_pred = y_pred+1
In [79]:
print(y_pred)
In [80]:
test['target'] = pd.DataFrame(y_pred)
In [81]:
test
Out[81]:
In [82]:
temp=pd.read_csv('/home/gaurav/Desktop/Social Cops/socialcops_challenge/socialcops_challenge/land_test.csv')
df=pd.DataFrame(data=test)
df['I6']=temp['I6']
target=df['target']
df=df.drop(['target'],axis=1)
df['target']=target
print(df)
df.to_csv('labelled_land_test.csv',index=False)